The LLM line of research will reach a capability plateau.

Drawn from what Gary Marcus said

Subjectsmeasuring intelligence LLMs scaling laws

What these subjects mean

measuring intelligence What it would mean to measure intelligence in a machine, as opposed to measuring how well it does one task it was trained for.

scaling laws Whether more compute keeps buying more capability, and where the curve now bends -- pre-training, post-training, test-time.

Mental modelS-curves and diminishing returns

What Gary Marcus actually said

Gary Marcus 7 Jun 2025 Cognitive scientist, professor emeritus of psychology… But this particular approach has limits that are clearer by the day. A knockout blow for LLMs?garymarcus.substack.com · 7 Jun 2025 All 6 korrents from this piece

Their wordsBut this particular approach has limits that are clearer by the day.

Where this was said

↗A knockout blow for LLMs?garymarcus.substack.com

7 Jun 2025

All 6 korrents from this piecethis one is 6th

No longer holds this

François Chollet 7 Aug 2026, 18:34 UTC Creator of the Keras deep-learning library and the ARC-AGI benchmark In the era of base LLM scaling (2022-2024), I believed the LLM line of research would reach a capability plateau (as later seen with base LLMs). In late 2024, after the o3 test-time compute demo, I changed my views: the new models were showing genuine fluid intelligence, and with this new line of work, the LLM line of research could achieve unbounded capability scaling. "There will be no wall." @fchollet on Xx.com · 7 Aug 2026

Recalls holdingIn the era of base LLM scaling (2022-2024), I believed the LLM line of research would reach a capability plateau (as later seen with base LLMs). In late 2024, after the o3 test-time compute demo, I changed my views: the new models were showing genuine fluid intelligence, and with this new line of work, the LLM line of research could achieve unbounded capability scaling. "There will be no wall."

↗@fchollet on Xx.com
7 Aug 2026

Added to korrents 29 Aug 2026 · How quotes work · Something wrong? Tell us

Holders over time

Do you hold this korrent?Do you also believe this?

I hold this too

Where this collides

Cannot both be true

AI model capabilities have not slowed down so far, and their progress probably won't stop anytime soon.

Chollet says the LLM line of research reaches a capability plateau; Ronacher says model capability progress is not going to slow down soon. A plateau and a continuing climb are the same curve described two incompatible ways.
In tension with

AI progress is faster than people expect, and ordinary scaling can be enough to solve problems that looked very hard.

Alexander took his image-model bet as evidence that scale and ordinary progress can be enough to solve problems that looked hard. Chollet says the LLM line of research reaches a capability plateau and that new architectural ideas are what further progress needs. One reads the last few years as the trend continuing; the other as the run-up to it stopping.

Related korrents

Shared subject first, then similar wording. They may agree or disagree.

On the map

Loading the map… or open it on its own page

Open the map on its own page